Last updated: November 16, 2026
goose 1.59, the open-source AI agent from the Agentic AI Foundation, shipped November 10, 2026. It lets live voice use any GPT-Live compatible server and custom instructions, resumes finished subagents with follow-ups, adds opt-in deferred MCP tool schemas and ACP session tags, and gives developers stop, approval and tool hooks.
Contents
- What goose is and what changed
- Configurable live voice
- Resumable subagents
- Deferred MCP tools and session tags
- Hooks in the goose Development Kit
- goose 1.59 compared with 1.58
- How to install or update
- Limitations
- Frequently asked questions
What goose is and what changed
goose is a general-purpose AI agent that runs on your own machine, with a desktop app for macOS, Linux and Windows, a CLI and an API. It is written in Rust, licensed under Apache 2.0, works with more than 15 model providers, and is part of the Agentic AI Foundation at the Linux Foundation, according to the goose documentation.
Version 1.59.0 was published on GitHub on November 10, 2026. Two patch releases followed: 1.59.1 on November 14 and 1.59.2 on November 16, which fixes how the Databricks provider handles thinking blocks. The 1.59.0 release notes list these features:
| Feature | Area | What it does |
|---|---|---|
| Configurable GPT-Live endpoint, model and key | Live voice | Points voice sessions at any GPT-Live compatible server |
| Configurable and per-start voice instructions | Live voice | Adds a name, speaking style or agent context to the voice model |
| Resume completed delegates with a follow-up | Subagents | Continues a finished subagent with its earlier context |
| Minimal deferred MCP tools | Tools | Sends tool names first and loads schemas on request |
| Filter session/list by client tags | ACP | Finds tagged sessions through an index lookup |
| Stop, approval and tool hook operations | GDK | Reusable hooks for agents built on goose-agent |
Configurable live voice
Before 1.59.0, live voice used api.openai.com. Pull request 13360 adds three settings:
GOOSE_LIVE_VOICE_ENDPOINT: the full sessions URL. When unset or blank, goose keepshttps://api.openai.com/v1/live/sessions.GOOSE_LIVE_VOICE_MODEL: defaults togpt-live-1.GOOSE_LIVE_VOICE_API_KEY: a secret used only with a custom endpoint. A custom endpoint never receivesOPENAI_API_KEY.
The endpoint must be http or https with a host, and plain http is allowed for servers on a local network. If the endpoint value cannot be read, voice becomes unavailable instead of falling back to OpenAI.
Pull request 13375 adds instructions in two layers. GOOSE_LIVE_VOICE_INSTRUCTIONS applies to every voice session on that install, and an optional instructions field on the live voice start request applies to one session. goose always sends its built-in instructions first, then the configured text, then the per-start text.
Resumable subagents
goose subagents are independent instances that take on a task, such as a code review or research, while your main conversation stays focused. The subagents guide says goose can start them on its own in autonomous mode, the default, or you can ask for them in plain language.
Before 1.59.0, a request to resume a completed delegate was rejected as already completed. Per pull request 13106, a finished delegate can now be resumed with a follow-up and keeps its earlier context. A resume must include new instructions.
A practical pattern:
- Ask goose to use a subagent for a bounded job, for example “Use a code reviewer subagent to check the auth module for security issues.”
- Read the subagent’s findings in the main session.
- Ask goose to send that same reviewer a follow-up, such as “Now check the fixes you suggested against the session code,” so it reuses what it already read instead of starting cold.
Other coding agents have moved the same way. Our guides to OpenCode 2.3 background subagents and Claude Code Projects with /btw and /fork cover similar features.
Deferred MCP tools and session tags
goose publishes every MCP tool schema to the model by default, which uses context when many extensions are on. Pull request 13378 adds an opt-in alternative:
- Set
GOOSE_TOOL_DEFERRAL=lookupas an environment variable or in config.yaml. Lookup is the only strategy implemented so far. - Optionally list tools that should stay fully loaded, for example
GOOSE_TOOL_DEFERRAL_ALWAYS_LOAD=serverA__toolA,serverB__toolC. - Start a session. goose lists the available tool names in the system prompt, and the model calls
enable_toolswith the names it needs, which publishes those schemas on its next response.
For client developers, pull request 13173 lets an ACP client tag a session when it creates it, with up to 16 string tags, and then list only sessions that carry every given tag. This matters if you drive goose from an editor through ACP; see our Zed Agent guide for one such client.
Hooks in the goose Development Kit
The goose Development Kit (GDK) exposes the runtime behind goose for building your own agents. Pull request 13169 adds reusable operations to goose-agent: a stop hook that can accept a final answer or send the model back to work, tool approval, a hook that runs just before each approved tool call, and one that runs right after each call.
goose 1.59 compared with 1.58
Version 1.58.0, released November 9, focused on context visibility. The 1.58.0 release notes and linked pull requests describe these changes:
| Release | Main additions |
|---|---|
| 1.58.0 (Nov 9) | Context window report and viewer with a Compact now button in the desktop app; opt-in, observation-only context relevance operation; AIHubMix, RouterPlex and Lemonade providers; thinking collapsed by default in the desktop app; session ID in /status |
| 1.59.0 (Nov 10) | Configurable live voice, resumable subagents, deferred MCP tools, ACP session tags, GDK hooks |
The context window viewer breaks down the system prompt, extension instructions, tool definitions and conversation. Its counts are local estimates, so they will not match the provider-billed total exactly.
How to install or update
The installation guide gives these routes for the CLI:
- On macOS or Linux, run
curl -fsSL https://github.com/aaif-goose/goose/releases/download/stable/download_cli.sh | bash. AddCONFIGURE=falsebeforebashto skip interactive setup. - On macOS with Homebrew, run
brew install block-goose-cli, orbrew install --cask block-goosefor the desktop app. - To update a script install, run
goose update.
If you run local models, our guide to Ollama 0.40.15 covers a common provider for goose.
Limitations
- Unstable APIs. The live voice start method sits under an unstable ACP namespace, so client integrations may need changes later.
- Opt-in features. Tool deferral and context relevance are off by default, and context relevance only records judgments without changing what the model sees.
- Mode limits. Subagents do not run in manual approval, smart approval or chat-only modes, and time out after five minutes by default.
Frequently asked questions
What is new in goose 1.59?
goose 1.59.0, released November 10, 2026, adds configurable live voice (custom GPT-Live compatible endpoint, model, key and extra instructions), resumable completed subagent delegates, opt-in deferred MCP tool schemas, tag filtering for ACP sessions, and stop, approval and tool hooks in the goose Development Kit.
How do I update goose to the latest version?
If you installed the CLI with the download script, run goose update. If you used Homebrew, Homebrew can handle updates for the block-goose-cli formula or the block-goose desktop app cask.
Can goose live voice use a server other than OpenAI?
Yes. Since 1.59.0 you can set GOOSE_LIVE_VOICE_ENDPOINT to any GPT-Live compatible sessions URL, GOOSE_LIVE_VOICE_MODEL to the model (default gpt-live-1) and GOOSE_LIVE_VOICE_API_KEY for that endpoint. A custom endpoint never receives your OPENAI_API_KEY.
Can I continue a goose subagent after it finishes?
Yes. goose 1.59.0 lets a completed delegate be resumed with a follow-up that keeps its earlier context. Earlier versions rejected the request as already completed. A resume must include new instructions.
What does GOOSE_TOOL_DEFERRAL do?
Setting GOOSE_TOOL_DEFERRAL=lookup makes goose list MCP tool names in the system prompt instead of sending every tool schema. The model calls enable_tools with the names it needs. Deferral is off when the variable is unset.
Which modes allow goose subagents?
The goose documentation says subagents run automatically in autonomous permission mode, which is the default. They are disabled in manual approval, smart approval and chat-only modes, and a subagent times out after five minutes by default.
Sources
- goose v1.59.0 release notes, GitHub (November 10, 2026)
- goose v1.58.0 release notes, GitHub (November 9, 2026)
- Subagents, goose documentation
- Install goose, goose documentation
Featured image: Canada goose in flight, photo by Alan D. Wilson, modified by Diliff, via Wikimedia Commons, CC BY-SA 2.5. The photo is illustrative and is not a goose product logo.